HyperTransport Dennis Vega Ryan Rawlins What is HyperTransport (HT)? • A point‐to‐point interconnect technology that links processors to other processors, coprocessors, I/O controllers, and peripheral controllers. How Does HyperTransport Work? • Point‐to‐Point Link Topology – Structure of the link • Electrical/Signal Characteristics – Physical Characteristics of the signal interface • Packet Protocols – How data is organized and transferred across the link Point‐to‐Point Link Topology • Employs dual, unidirectional data links with a concise signal set using 1.2V LVDS • The daisy‐chain topology includes a required “host” device and at least one end‐point or “cave”. • Optional: – “Tunnel” devices that connect the link to other HT devices – “Bridge” devices that interface with non‐HT interconnect technologies. Point‐to‐Point Link Topology Electrical/Signal Characteristics • The unidirectional links include a data path, one or more clock lines, and a single control line. Electrical/Signal Characteristics • Each LVDS line is implemented by means of a balanced or differential line – Prevents electrical noise from affecting and potentially corrupting the signal detection process • The HT architectures can be implemented with fewer PCB traces and less signal shielding precautions due to the lower number of signal lines for a given bandwidth. Electrical/Signal Characteristics Packet Protocols • Command information is transferred as a control packet of 4 or 8 bytes. • HT data traffic is delivered as a data packet that consists of an 8‐ or 12‐ byte header followed by a 4‐64 byte data payload. • For a write request, there is an 8‐byte control packet followed by the data packet. • For a read request, there is an 8‐byte control packet, followed by a 4‐bytes read response packet, and followed by the data packet. Packet Protocols Packet Protocols • PRI attaches a high priority request command, consisting of 8‐bytes, within a possibly long but lower priority data transfer. Packet Protocols • Commands and Data are separated into one of three types of virtual channels: – Non‐posted requests – require a response from the receiver (all read and some write requests) – Posted requests –do not require a response from the receiver (write requests) – Responses – replies to non‐posted requests (read responses or target done responses to non‐posted writes) History • Introduced April 2, 2001 by the HyperTransport Consortium • The Consortium consisted of several initial companies including IBM, Apple and AMD to develop a low latency, royalty free I/O bus design that was high performance and fully scalable History • Influential factors: the need for PCI compatibility and a requirement for low cost implementation – Conformed to all ordering and enumeration properties of the PCS standard – Low cost implementation achieved with the point‐ to‐point link structure and the 1.2V LVDS Evolution • Initially started with version 1.0 in 2001 and 1.1 in 2002 – Max freq. of 800MHz – Max bandwidth at 16‐bits of 3.2 GB/s – Max bandwidth at 32‐bits of 6.4 GB/s – Max aggregate (bi‐directional) bandwidth of 12.8 GB/s – DirectPackets Data Streaming (1.1 only) • Version 2.0 2004 – Max freq. of 1.4 GHz – Max bandwidth at 16‐bits of 5.6 GB/s – Max bandwidth at 32‐bits of 11.2 GB/s – Max aggregate (bi‐directional) bandwidth of 22.4 GB/s – DirectPackets Datastreaming – PCI Express Mapping Evolution • Version 3.0 in 2006 – Max freq. of 2.6 GHz – Max bandwidth at 16‐bits of 10.4 GB/s – Max bandwidth at 32‐bits of 20.8 GB/s – Max aggregate (bi‐directional) bandwidth of 41.6 GB/s – AC Operation: Capacitive coupling with AC/DC autosensing and autoconfiguration – Link splitting – Hot‐plugging – Dynamic link clock/width adjustment – DirectPackets Datastreaming – PCI Express Mapping • Version 3.1 2008 – Max freq. of 3.2 GHz – Max bandwidth at 16‐bits of 12.8 GB/s – Max bandwidth at 32‐bits of 25.6 GB/s – Max aggregate (bi‐directional) bandwidth of 51.2 GB/s – Same features as 3.0 Evolution • Currently at Version 3.1 • Updates approximately every 2 years • Last revision was 3.10c posted by the Consortium on 5/6/2010 • Small and minor updates in revisions • The Consortium has stated that its mission is to keep HyperTransport specifications “comfortably” ahead of industry requirements • HyperShare is also a future concept utilizing the best of interconnect fabric technologies such as HyperTransport, Ethernet and PCI Express Advantages • Industry’s lowest latency, highest bandwidth bus for maximum performance • Processor native (embedded in processor chip) to remove latency from processor to I/O interface control logic found in other standards such as PCI Express • Native to several companies including AMD, Nvidia and Bay Microsystems • Low cost • Royalty free • Application flexibility (processor to processor, processor to peripheral or I/O and board to board) and excellent scalability • Low Power consumption • Backwards compatibility amongst all older versions and transparent extensibility to other standards such as PCI Disadvantages • No dedicated I/O address space unlike a PCI bus • No dedicated configuration address space also unlike a PCI bus • CPU I/O space is memory‐mapped to a high address range • Addresses must be reserved for I/O and can’t be used for physical storage –may be temporary or permanent • Two unidirectional LVDS lines are required for every signal line Applications • Many companies utilize HyperTransport in their systems including IBM and AMD • Used in many types of systems including personal computers, super computers and game consoles • Also used for front‐side bus replacement, multiprocessor interconnections, router bus replacement, co‐processor interconnections and add‐on card connections Applications • Two types of HyperTransport cores exist including the standard core and the coherent core • Coherent core provides similar functionality as the standard core with the ability to connect to coherent devices such as coherent cache (consistent cache blocks in multiprocessing systems) Conclusion • Low cost, low latency bus mostly used in front‐side bus replacement • Full backwards compatibility • Wide array of uses • Memory‐mapped I/O • Dual unidirectional data links • Managed by the HyperTransport Consortium Questions? References • General information: www.hypertransport.org • Overview Information: http://www.hardwaresecrets.com/article/19 • HT‐Core: http://ra.ziti.uni‐heidelberg.de/coeht/?page=projects&id=htcore • cHT‐Core: http://ra.ziti.uni‐heidelberg.de/coeht/?page=projects&id=chtcore • Reference Designs: http://www.hypertransport.org/default.cfm?page=ReferenceDesigns • Connectors: http://www.hypertransport.org/default.cfm?page=HTConnectorsAndCables • Documentation: http://www.hypertransport.org/default.cfm?page=TechnologyWhitePapers • Wikipedia: http://en.wikipedia.org/wiki/HyperTransport • HyperShare: http://www.marketwire.com/press‐release/hypertransport‐consortium‐discusses‐benefits‐ combining‐hypershare‐technology‐with‐ethernet‐1399316.htm • http://www.reuters.com/article/2011/02/22/idUS135038+22‐Feb‐2011+MW20110222 • http://www.hpcwire.com/hpcwire/2010‐12‐01/hypertransport_consortium_unveils_hypershare_platform.html.